From be61384272907bc4882c4d09d427bad8f8e3dd28 Mon Sep 17 00:00:00 2001
From: amcamd <andrew.chapman@amd.com>
Date: Tue, 4 Aug 2020 12:29:08 -0500
Subject: [PATCH] cherry-pick Remove workaround for K==0 from rocBLAS-internal
 commit fc9729ce50366c5da2d547f1c0321458129c6152

---
 clients/gtest/gemm_gtest.yaml |  5 +++--
 library/src/tensile_host.cpp  | 19 +++++--------------
 tensile_tag.txt               |  2 +-
 3 files changed, 9 insertions(+), 17 deletions(-)

diff --git a/clients/gtest/gemm_gtest.yaml b/clients/gtest/gemm_gtest.yaml
index 4a72630af..479884ff3 100644
--- a/clients/gtest/gemm_gtest.yaml
+++ b/clients/gtest/gemm_gtest.yaml
@@ -2911,9 +2911,10 @@ Tests:
   alpha_beta: *alpha_beta_range
   K: 0
   matrix_size:
-    - { M:  1,   N:  2 }
-    - { M:  3,   N:  5 }
+    - { M:  1,   N:    2 }
+    - { M:  3,   N:    5 }
     - { M:  512, N:  100 }
     - { M:  63,  N:  512 }
+    - { M: 100,  N: 1000 }
 
 ...
diff --git a/library/src/tensile_host.cpp b/library/src/tensile_host.cpp
index 3469ac643..be8acfaf1 100644
--- a/library/src/tensile_host.cpp
+++ b/library/src/tensile_host.cpp
@@ -140,11 +140,6 @@ namespace
         freeIndex[0].c = freeIndex[0].d = 0;
         freeIndex[1].c = freeIndex[1].d = 1;
 
-        // Tensile does not support 0-sized dimensions. For when k == 0, we still need to
-        // multiply C by beta, but not add any of the rank-0 dot products. As a workaround,
-        // we pass k = 1 and set alpha == 0, since alpha == 0 has the same effect as k == 0.
-        auto k = prob.k == 0 ? 1 : prob.k;
-
         // clang-format off
 
         // If A is transposed, swap the free and bound dimensions and their ranks
@@ -152,7 +147,7 @@ namespace
         {
             a = {
                     Tensile_Ti,
-                    {k, prob.m, prob.batch_count},
+                    {prob.k, prob.m, prob.batch_count},
                     {prob.row_stride_a, prob.col_stride_a, prob.batch_stride_a}
                 };
             freeIndex[0].i  = 1;
@@ -162,7 +157,7 @@ namespace
         {
             a = {
                     Tensile_Ti,
-                    {prob.m, k, prob.batch_count},
+                    {prob.m, prob.k, prob.batch_count},
                     {prob.row_stride_a, prob.col_stride_a, prob.batch_stride_a}
                 };
             freeIndex[0].i  = 0;
@@ -178,7 +173,7 @@ namespace
         {
             b = {
                     Tensile_Ti,
-                    {prob.n, k, prob.batch_count},
+                    {prob.n, prob.k, prob.batch_count},
                     {prob.row_stride_b, prob.col_stride_b, prob.batch_stride_b}
                 };
             freeIndex[1].i  = 0;
@@ -188,7 +183,7 @@ namespace
         {
             b = {
                     Tensile_Ti,
-                    {k, prob.n, prob.batch_count},
+                    {prob.k, prob.n, prob.batch_count},
                     {prob.row_stride_b, prob.col_stride_b, prob.batch_stride_b}
                 };
             freeIndex[1].i  = 1;
@@ -303,11 +298,7 @@ namespace
 
         // alpha and beta are stored by value in Tensile::TypedContractionInputs
         // alpha and beta are copied from host to Tensile::TypedContractionInputs
-        // We set alpha = 0 if k == 0 (see above)
-        if(prob.k == 0)
-            memset(&inputs.alpha, 0, sizeof(inputs.alpha));
-        else
-            AlphaBeta<Ti, To, Tc>::copy(&inputs.alpha, prob.alpha);
+        AlphaBeta<Ti, To, Tc>::copy(&inputs.alpha, prob.alpha);
         AlphaBeta<Ti, To, Tc>::copy(&inputs.beta, prob.beta);
 
         return inputs;
diff --git a/tensile_tag.txt b/tensile_tag.txt
index 854707180..316ecaff3 100644
--- a/tensile_tag.txt
+++ b/tensile_tag.txt
@@ -1 +1 @@
-19319869e83243c5e3ca649532d2951de0ba35be
+af71ea890a893e647bf2cf4571a90297d65689ca